Cost controls

Signals is the highest-cost feature per run (web_search + LLM + retries). Xura has four defensive layers.

​

1. 24-hour cooldown

Once a run produces signals, Refresh Research locks for 24 hours. Both client-side (button disabled with countdown) and server-side (returns 429 if bypassed). Empty runs don’t count — you can retry immediately if nothing was found.

​

2. Web-search cap

The web_search tool is capped at 5 uses per run. Each search costs (fetches + processes a page); more than 5 rarely improves signal quality.

​

3. Output cap

max_tokens: 2500 limits how much prose the LLM can emit. 6-12 signal JSON + a 60-120 word digest fits comfortably.

​

4. Prompt caching

The fixed instruction block (schema + rules) is marked cache_control: ephemeral. Anthropic serves it from cache on subsequent runs at ~10% of the input cost.

​

Typical cost per run

Depends on your model choice: Numbers assume ~4 web_search calls, 2500 output tokens, cached system prompt after the first run.

​

Where to see actual spend

AI Studio → Activity log shows every LLM call with token counts + source. Filter by source: market to see Signals-only usage.